Period Prediction of Sinhala Epigraphical Scripts using Convolutional Neural Networks

Inscriptions are important resources to know our history. The study of recognizing epigraphical scripts is a challenging task since the shapes of the characters were changed over the time and different sets of characters were used in different eras. Period prediction of epigraphical scripts is an important initial step in automated inscription character recognition systems, and also it helps archaeologists to find the era of an inscription in real-time. In this paper, we propose a novel approach to classify the era of Sinhala epigraphical scripts into five different periods using the images of Sri Lankan inscriptions. Since no previous studies were conducted, we have constructed a dataset for Sinhala ancient characters. Deep transfer learning techniques are utilized to train a CNN model with fewer numbers of samples. Moreover, a channel attention module is included to boost the character-wise features in classification. Based on the experimental results, the proposed approach showed 90.60% of average classification accuracy.